Mortality Risk Prediction Of Acute Coronary Syndrome Based On Deep Learning

Acute coronary syndrome(ACS)is a common and serious cardiovascular disease.Patients with ACS often present with symptoms such as paroxysmal chest pain and tightness in the chest,which can lead to arrhythmia,heart failure and even sudden death.ACS has the features of high morbidity and high mortality,and is also the main cause of death worldwide,which seriously threatens people’s health and life safety.Due to the rapid onset and rapid development of ACS,every second counts in the rescue,if the risk of death of ACS patients can be predicted,high-risk groups can be identified as early as possible,helping doctors to judge the progress of the patient’s condition,and then assisting doctors in making timely treatment decisions and reducing the risk of death.The risk of death of patients is of great significance for the clinical care of patients and improving the survival rate of patients.There are three problems in the death risk prediction models widely used in clinical practice at present: First,ACS is a clinical syndrome,in the diagnosis process,not only ECG and myocardial enzymology tests should be performed,but also coronary angiography and other examinations should be further performed when necessary.The examination data has the problem of high dimension and large redundancy,and direct input into the model will cause "dimension disaster" and affect the accuracy of the experimental results;Secondly,most of them are based on statistical methods for modeling,taking specific clinical examination indicators as risk features as model input,and using logistic regression method for prediction.These methods can better pay attention to the statistical correlation in the data,but the real clinical ACS patient data have the problem of unbalanced data feature distribution,such as the elderly population accounts for a large proportion,the individual feature distribution of patients will have certain bias,the model will better fit the feature distribution law of this population in the learning process,and cannot be generalized to the prediction task of other age groups;Third,clinical studies have shown that the introduction of specific treatment indicators into risk features is helpful to improve the accuracy of ACS risk prediction.However,most models at present only directly splice the data of treatment indicators and examination indicators into risk features input,and do not well integrate different features in modeling,and lack of mining their correlation.To address the above issues,this paper investigates the prediction of ACS mortality risk using deep learning techniques by analyzing clinical data of ACS patients provided by a tertiary hospital in Shanghai,with the following three main components:(1)To solve the problem of redundancy and noise in high-dimensional data,a method of selecting risk features of ACS patients based on objective weight is proposed.Firstly,performing feature importance scoring on risk features of the ACS patient by using three feature selection methods of principal component analysis,random forest and Shapley weighted interpretable value.Then,using the volatility of different scores under the features and the correlation between the features as objective weight values for measuring the features,and performing feature importance ranking again based on the values,so that the influence of invalid features is eliminated,and the features which are efficient and highly suitable for a target task are screened out.(2)Aiming at the unbalanced distribution of features in real data,An ACS mortality risk prediction method based on causal reweighting is proposed.The tab Transformer method is first used to extract the risk features and transform them into a context-embedded expression,and then the expression is reweighted by causal weight learning.The adjustment of the sample weight size enables the model to focus on the causal relationship between the features and the prediction target,eliminating the impact of imbalance caused by differences in the distribution of ACS clinical data,thus enhancing the stability and accuracy.By conducting experimental comparisons on data sets with different distributions,the results show that the method can effectively improve the stability of the model prediction.(3)A method for predicting the risk of death in ACS based on the interaction of attentional features is proposed to address the problem of feature fusion of treatment and examination indicator data.The method introduces treatment indicator data as risk features together with examination indicator data.Firstly,the treatment features are extracted from the patient’s treatment indicator data through a self-attentive mechanism,while the examination features are extracted from the patient’s examination indicator data through a causally weighted tab Transformer feature extractor;then the two are interactively fused using a factor decomposer to form patient interaction features,and the attention weights of these interaction features are further learned to obtain the weighted patient interaction features,so as to improve the attention to important interaction features;finally,the weighted interaction feature information is input into the model for training.Finally,the weighted interactive feature information is used for prediction.Through experimental comparison with other feature fusion methods,it is verified that this method can improve the accuracy of model prediction.